MetaPatchAI home
AI

"LLM Voice": Why Readers Bail on AI-Sounding Writing, and 8 Ways to Fix Your Drafts

Source: martinfowler.com, Martin Fowler

Compiled by MetaPatchAI · Published , 12:34 AM ET

MetaPatchAI header for ""LLM Voice": Why Readers Bail on AI-Sounding Writing, and 8 Ways to Fix Your Drafts": headline on a dark background with abstract neon-green network-node art.

Martin Fowler, the software author, put a name to something many readers feel: "LLM Voice." It's the sense, partway through a post, that you're reading machine output, and that your wish to keep reading is fading.

What Fowler argues (summary)

  • The reaction is human and fast. By the time you'd run a detector tool, the reader has already lost interest, so detector scores miss the point.
  • He doesn't trust AI-text detectors' accuracy. He also doubts his own radar, because some phrases that annoy him are simply generational or regional ways of speaking, not machine tells.
  • He jokes that the giveaways are not just em-dashes but "load-bearing tapestries", the overworked metaphors and filler phrasing of model prose.

Read the full piece. It's short: https://martinfowler.com/bliki/LLM-Voice.html

8 practical fixes (our checklist, not Fowler's)

  1. Lead with a specific fact, number or scene, not a framing sentence ("In today's fast-paced world…").
  2. Cut the summary echo. Models often restate the point at the end of each paragraph. Delete it.
  3. Kill the stock metaphors: tapestry, landscape, realm, delve, game-changer, unlock.
  4. Break the rule of three. Models love lists of three adjectives. Use one good one.
  5. Add first-hand material: something you tested, a screenshot, a mistake you made. A model can't fake this, and readers notice.
  6. Vary sentence length on purpose. Short. Then a longer one that carries real detail.
  7. Take a position. "It depends" endings are a tell. Say what you'd do.
  8. Read it aloud. If you wouldn't say it to a friend, rewrite it.

Why this matters for anyone publishing with AI help

Search engines and readers both reward pages that show first-hand experience. AI is fine for research and first drafts. The parts that make a page worth reading (testing, judgment and voice) still have to come from a human.

Video

Video by BlueFx Video AI on YouTube.

Video by Craig Does AI on YouTube.

Sources

What is confirmed, and what is not

  • ✅ The summary of Fowler's points comes from the essay text read on 2026-10-29. The quoted fragment ("load-bearing tapestries") is short and attributed.
  • The 8 fixes are MetaPatchAI's own editorial advice and are labelled as such. They make no claims about how Google ranks pages.

Related

Up: AI